Proof
Every figure on this page was computed by a query when you loaded it. None of them is typed into the HTML. Each one says what it counts, over what window, and as at when. Where a query did not answer, it says so rather than showing you the last number we happened to know. A zero is printed as a zero.
Generated 21 Aug 2026, 23:14 UTC. Reload and the stamps move.
Kate sits a fixed exam against an invented Leeds bakery, and a separate model, deliberately not the one being tested, marks every answer. We wrote the questions, so this is a self-set exam and we say so. What it can still prove is regression: the number moves the morning after a deploy that made her worse, and it is published whether it went up or down.
Nightly consultant-exam pass rate
100%
52 of 52 questions in the frozen panel produced an answer the judge could score. A pass means a separate model (not the one being tested) rated the answer worth paying a consultant for.
Nightly runs kept on record
38
One row per night, written by the 03:00 job. First 12 Jul 2026, latest 21 Aug 2026.
Weekly benchmark mean score
8.7 / 10
Mean of the judge's 0-10 scores across 50 of 50 questions, tiered easy to hard, all answered against one invented bakery.
MarketingBench pass rate, public tasks
100%
9 of 9 scored public tasks passed, out of 13 published tasks. 3 further tasks are held back so the public list cannot be overfit, and are excluded from this figure.
Questions in the panels the scores above come from
115
52 in the frozen nightly exam, 50 in the weekly benchmark, and 13 published MarketingBench tasks you can read in full at /marketingbench. Every question is written by us and answered against an invented business, never a customer's own data. No human consultant was scored alongside.
An AI marketing department that only works when someone opens a tab is not a hire, it is a tool. So here is our own machine's attendance record: every scheduled job this deployment declares, judged against its own tolerated silence, computed when you loaded this page. A job that has never run and has not been watched long enough to judge is counted separately, as neither running nor broken, because alarming on ignorance is how a signal gets ignored.
Scheduled jobs registered
21
Every scheduled job this deployment declares, none excluded. A job cannot be added, removed or re-timed in vercel.json without this registry moving with it, because a test compares the two lists both ways.
Jobs that completed a run inside their own window
20 of 21
Counted from the completion stamp each job writes as its last statement, so a job that starts and dies does not count here.
Everything above is about us: an exam Kate sits against an invented bakery, and our own machine’s attendance record. Nothing above is derived from a customer’s workspace, because we are a processor of that content and the purposes we published do not include making public statistics out of it. Anonymising it would itself be processing, so there is no aggregate small enough to slip through the gate. You will not find a “businesses served” counter here either. A headcount is not evidence that the work gets done, and when we do publish one it will be on this page with its query beside it, like everything else.
Signed permissions on file to publish a named customer's numbers
0
None. So there are no case studies on this site, and /customers is empty rather than populated with invented ones. When this moves, the case study appears and this figure moves with it.
Minimum businesses in any cell before we would publish a breakdown
20
Across the 8 business types we offer, and worse once you cross them with region, this threshold suppresses every cell we could build today. We do not compute the cell sizes: the lawful-basis gate below fails first, and running the query anyway would itself be the processing we say we are not doing.
Jobs not running
0
0 invoked and dying part-way through; 0 not being invoked at all. Which jobs, and what stops happening while they are quiet, is on our internal health endpoint rather than here.
Jobs we cannot yet judge
1
Never stamped, and we have not been watching for longer than the job's own interval, so silence is not yet evidence. Reported as its own figure rather than counted as running or as broken.
Posts published for customers, by channel, over the last 30 days
Suppressed
We hold this and could total it in one query. We are a processor of workspace content and our published purposes are exhaustive: authenticating you, generating drafts, routing approvals, producing reports, supporting you. Deriving a public statistic is a new purpose, and anonymising it is itself processing, so an aggregate does not escape the gate. The fix is a contract change we have not made, not a bigger sample.
Businesses using Marketeer, broken down by trade and region
Suppressed
Trade plus region plus a date is a quasi-identifier at any size, and at ours it is a name. Even if the basis existed, every cell would fall under the minimum above and be suppressed, which is why you would be reading a table of blanks rather than a benchmark.
Measured reach or engagement uplift after one of Kate's recommendations
Suppressed
These figures are the business's own trading performance and are confidential to them under our own terms. They are also platform-derived, and what we may do with platform insights beyond serving the customer who provided them is not ours to decide.
A named business, with the numbers we measured together
Suppressed
This is the one place consent is the right instrument, because it is one named business agreeing to one named story. The wording is written and waiting. Until a signed one is on file, /customers stays empty. An empty page is not a weakness; a page of invented case studies would be.